| """SonoCNN demo: predict the bundled samples, a few held-out test images, |
| and — if the sibling ViewConsensus library is importable — an 8-view |
| consensus panel with a defer-to-human verdict on one test image. |
| |
| This is a mechanics demo on a 28x28 benchmark model, not a clinical tool. |
| """ |
| import os |
| import sys |
|
|
| import numpy as np |
|
|
| sys.path.insert(0, os.path.join(os.path.dirname(__file__), "..", "src")) |
|
|
| HERE = os.path.dirname(os.path.abspath(__file__)) |
| PKG = os.path.dirname(HERE) |
|
|
| from sonocnn import load_model, predict |
|
|
| model = load_model(os.path.join(PKG, "model.safetensors")) |
|
|
| print("== bundled demo samples (256x256 ultrasound stills, resized to 28x28) ==") |
| for name in ("benign", "malignant"): |
| r = predict(model, os.path.join(PKG, "assets", "demo_samples", f"{name}.png")) |
| print(f"{name}.png -> predicted {r['label']} " |
| f"(confidence {r['confidence']:.1%}; benign {r['probabilities']['benign']:.3f}, " |
| f"malignant {r['probabilities']['malignant']:.3f})") |
|
|
| print("\n== held-out BreastMNIST test images ==") |
| d = np.load(os.path.join(PKG, "assets", "breastmnist.npz")) |
| x_test = d["test_images"] |
| y_test = d["test_labels"].ravel() |
| labels = ("benign", "malignant") |
| for cls in (0, 1): |
| idx = int(np.where(y_test == cls)[0][0]) |
| r = predict(model, x_test[idx]) |
| print(f"test[{idx}] true={labels[cls]} -> predicted {r['label']} " |
| f"(confidence {r['confidence']:.1%})") |
|
|
| print("\n== ViewConsensus panel on the malignant test image (optional) ==") |
| try: |
| sys.path.insert(0, os.path.join(PKG, "..", "..", "libraries", "consensuskit", "src")) |
| from viewconsensus import consensus_from_probabilities |
|
|
| import torch |
| from sonocnn.infer import TRAIN_MEAN, TRAIN_STD |
|
|
| img = x_test[int(np.where(y_test == 1)[0][0])].astype(np.float32) / 255.0 |
|
|
| def shift(a, dy, dx): |
| return np.roll(np.roll(a, dy, axis=0), dx, axis=1) |
|
|
| views = [img, img[:, ::-1], img[::-1, :], shift(img, 0, 2), shift(img, 0, -2), |
| shift(img, 2, 0), shift(img, -2, 0), shift(img, 2, 2)[:, ::-1]] |
| batch = torch.from_numpy( |
| np.stack([((v - TRAIN_MEAN) / TRAIN_STD).astype(np.float32) for v in views]))[:, None] |
| with torch.no_grad(): |
| probs = torch.softmax(model(batch), 1).numpy() |
| res = consensus_from_probabilities(probs, labels=list(labels)) |
| verdict = getattr(res.verdict, "value", res.verdict) |
| print(f"consensus: {res.label} at {res.probability:.1%} " |
| f"(+/-{res.uncertainty:.1%} across 8 views) -> verdict: {verdict}, " |
| f"should_defer: {res.should_defer}") |
| except ImportError: |
| print("viewconsensus not importable here — install the sibling ViewConsensus " |
| "library to see the consensus panel.") |
|
|
